The Importance of Donor Variables on Pediatric Heart Transplant Survival

Author:
Sharff, Julia, Systems Engineering - School of Engineering and Applied Science, University of Virginia
Advisor:
Porter, Michael, DS-Academic Affairs Engineering Graduate, University of Virginia
Abstract:

Background: Waitlist mortality for pediatric heart transplant candidates remains high, with nearly 10% dying prior to receiving a heart. Despite this, there are a large number of donor hearts available for transplant, but refused due to poor donor quality. The 2020 ISHLT consensus statement suggested that few donor characteristics have an impact on post-transplant mortality. To test the validity of this, we utilized two machine learning models to evaluate the importance of various donor characteristics on patient survival post-transplant.
Methods: Random Forest (RF) and Lasso Logistic Regression (LR) were used to predict 1-year, 3-year, and 5-year post transplant survival using OPTN/UNOS data (2010-2019) for pediatric heart recipients. Candidates listed for multiple organs, re-transplantation, or with donors over the age of 30, were excluded from this study leaving a total of 3882 patients. RF and LR models were fit using combinations of donor, candidate, donor-candidate compatibility, and transplant predictor variables. A comparison of the AUC values, Brier scores, and log loss from 10-fold cross validation was used to assess differences in model performance.
Results: The LR models had higher average AUC, and lower brier score and log loss score, model performance for 1-year survival, with the best overall model (AUC = .754) coming from the candidate and donor variables. The random forest candidate model performed best for 3-year survival with an average AUC of .69. The RF models achieved better performance over LR for all 5-year survival models. The comparison of the model metrics from the different variable groups show that there is no statistically significant model improvement from the addition of non-candidate related variables. Further exploration into ischemic times suggest longer times are associated with reduced survival probability, making it difficult to determine its impact in models.
Discussion: The use of additional variables outside of echo and ischemic time when determining post-transplant success may be unnecessary. While ischemic time and echocardiograph measures were statistically insignificant as well, it is important to note that these conclusions are drawn based on contemporary donor heart selection practice, which displays little variability in accepted donor characteristics. Very few data points fell outside what is generally considered acceptable. The findings suggest that further evaluation on ischemic time and echo abnormality should be done to determine their impact on post-transplant survival.

Degree:
MS (Master of Science)
Keywords:
Pediatric Heart Transplantation, Machine Learning, Donor Characteristics
Language:
English
Rights:
All rights reserved (no additional license for public reuse)
Issued Date:
2024/04/21